LungIMPACT Explores AI Triage in Lung Cancer Detection

Immediate AI-based prioritization of chest x-rays did not accelerate the diagnostic pathway for suspected lung cancer in the LungIMPACT trial, according to results published in Nature Medicine by researchers from University College London Hospitals NHS Foundation Trust (UCLH), University College London (UCL), and the University of Nottingham.

AI prioritization was found to improve the speed of one step in the workflow—time from chest X-ray acquisition to report—but this did not appear to translate into faster progression through the broader diagnostic pathway. As reported by the investigators, referral rates, treatment start times, and stage at diagnosis were similar with and without AI prioritization.

“The bottleneck isn’t the reporting; it’s everything that happens next: telling the patient, the CT appointment, the clinic slot, the multidisciplinary meeting,” said principal investigator Nick Woznitza, MBE, PhD, FBIR, FCR, MASMIRT(AP), Consultant Radiographer at UCLH and Honorary Associate Professor of UCL, in a press release from the UCLH's Biomedical Research Centre. “We’ve shown that AI prioritization, by itself, cannot fix that.”

Study Details

In this prospective, multicenter, randomized controlled study, investigators evaluated the impact of AI-assisted prioritization of chest x-rays ordered in primary care on the lung cancer diagnostic pathway. Work list prioritization was randomized by day, with AI available in both study arms but used to expedite reporting only on intervention days.

A total of 93,326 chest x-rays from five NHS hospitals were included, with 45,987 and 47,339 in the prioritization “on” and “off” arms, respectively. All reporters received training on the use of the AI system prior to study initiation; the study did not alter the usual care pathway.

Time to CT and time to lung cancer diagnosis were evaluated as primary outcomes. The investigators identified the number of urgent suspected lung cancer referrals, incidence and stage of lung cancer, times to urgent referral and treatment, concordance between AI and radiology reports, and algorithm accuracy as secondary outcomes.

AI Methodology

The study used qXR (version 4.0; Qure.ai Technologies), a class IIb Conformité Européenne–certified deep learning algorithm designed for chest x-ray interpretation across 29 abnormality categories. This algorithm is already routinely used in clinical practice in some NHS Hospitals. The system functioned as a clinical decision support tool, analyzing pseudonymized images via a cloud-based platform and returning abnormality classifications with corresponding image annotations to the local Picture Archiving and Communication System.

On intervention days, cases classified as “qXR-suspected abnormal” generated active notifications within the radiology work list, enabling prioritization for immediate reporting.

The algorithm did not undergo updates during the study to ensure consistent performance.

Key Findings

Median time to CT did not differ between the groups with and without AI prioritization (both 53 days; ratio of geometric means = 0.97, 95% confidence interval [CI] = 0.93–1.02; P = .31). Among 13,347 CT scans identified, 2,766 were performed within 14 days of chest x-ray; in this subgroup, median time to CT was 8 days in both groups.

Lung cancer was diagnosed in 558 patients (0.6% of chest x-rays). Median time to diagnosis was 44 days with prioritization vs 46 days without (ratio of geometric means = 0.98, 95% CI = 0.83–1.16; P = .84).

The investigators reported no significant differences in time to lung cancer referral (14 vs 15 days; P = .13), time to treatment (76 vs 73 days; P = .99), or stage at diagnosis (P = .34).

AI prioritization did, however, reduce the time from chest x-ray acquisition to report, with median reporting times of 34 hours compared with 47 hours without prioritization.

Discordance between AI and radiologist findings occurred in 28,261 chest x-rays (30.3%). Actionable findings were identified in 6,750 of these cases (23.9%).

Time to diagnosis appeared to differ by concordance between AI and radiologist interpretation: 38 days when both identified abnormalities, 106 days in 53 cases where AI identified an abnormality not reported by the radiologist, and 177 days when both assessments were negative.

Commenting on these results, study investigator Arjun Nair, MD, Consultant Radiologist at UCLH and Honorary Associate Professor at UCL, said: “The cases where the AI spotted something the radiologist did not flag are the ones that interest us most. These patients waited much longer for a diagnosis. We need to understand whether there is a pattern, as this has implications not just for how we use the AI to spot these cases earlier but also potentially how we train radiologists and radiographers of the future.”

The National Optimal Lung Cancer Pathway specifies CT within 72 hours of a suspicious chest x-ray—ideally on the same day; in this study, 172 CT scans were performed on the same day and 477 within 72 hours.

The investigators concluded, “AI prioritization of chest x-rays requested by UK primary care has no significant impact on the lung cancer pathway. Therefore, chest x-ray AI deployments should not include work list prioritization in this context. Future research should differentiate between primary pathway changes and the direct impact of AI.”

In the news release, chief investigator and corresponding study author David R. Baldwin, MD, FRCP, Professor of Medicine at the University of Nottingham and Consultant Respiratory Physician at Nottingham University Hospitals NHS Trust, added: “The technology is maturing, we need to understand how busy departments can benefit most given the challenges they face. And the pathway is not ready for this. If we want AI to make a real difference to lung cancer outcomes, we need to redesign how the NHS responds when the AI raises an alert, and that means better coordination, more resource, and a genuine commitment to same-day action.”

DISCLOSURES: For full disclosures of the study authors, as well as funding information and code from the statistical analysis, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.